System design for physical AI: leak detection, compressor health and OT security evidence for a US gas transmission estate
This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/pipeline-integrity-gas-transmission-us/ . The operator is described by class, never by name. A gas transmission operator in the United States runs more than 3,000 miles of pipe, more than a dozen compressor stations,…
This paper outlines a system design for physical AI in the US gas transmission sector. The operator is classified, not named. The system monitors 3,000 miles of pipe, over a dozen compressor stations, several control centers, and more than 100 compressor units. Four critical inquiries are made hourly: determining if pressure and flow deviations indicate leaks or line pack replenishment, whether compressor units are failing, identifying unauthorized individuals in the OT network of a station, and verifying the availability of security directive evidence.
System design ensures no interaction with field devices. Control authority remains within certified pathways. Control room assistants communicate alerts to operators in plain language, drafting responses. Pipeline controllers make final decisions, adding no additional alarms.
Data streams from stations through passive taps and SPAN ports, avoiding active scanning that could disrupt legacy devices. Each station exits its OT zone via a hardware enforced one-way data diode, chosen over bidirectional firewalls due to a major US pipeline shutdown caused by IT/OT coupling failure. Model and container updates enter through a separate, approved path, bypassing the data diode.
Four data sources enter read-only through three adapter families: Telemetry, Ops systems, and Integration. Each event carries a provenance stamp, and a unified object model serves as the sole writing layer. This model is backed by Apache Kafka 4.3, InfluxDB 3 Core, MediaMTX, K3s, CVAT, linuxptp, Keycloak, and Prometheus with Grafana and Loki.
The ontology encompasses compressor stations, meter stations, pipeline segments, compressor units, OT devices, pipeline controllers, alarms, leak or abnormal operation events, remote access sessions, work orders, security directive control, historian, CMMS, and hazardous area classification. These entities are linked through 13 typed relationships. Queries trace methane readings to hazardous area classifications, alarm confirmations, work orders, and regulatory references.
Models are built directly from data through Chronos-2 for pressure, flow, vibration, and temperature forecasting, and RF-DETR for camera monitoring of gas plumes, vent events, hot spots, people, and service trucks. These models run on edge modules consuming minimal power while retaining data through backhaul loss.
The entire system is underpinned by Apache-2.0 licensed models, Mistral Large 3 675B for complex reasoning and interpretation, and BGE-M3 for video analysis. The station hardware cost is omitted, as the focus is on the AI system design. Phase 1 involves testing at 16 pilot stations, limited to API RP 1130 gate. All models and systems run in shadow before any controller-facing alerts are issued, ensuring rigorous validation before implementation.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
